Showing posts with label continual learning. Show all posts
Showing posts with label continual learning. Show all posts

Daily Tech Digest - July 16, 2024

Learning cloud cost management the hard way

The rapid adoption of cloud technologies has outpaced the development of requisite skills within many organizations, leading to inefficiencies in provisioning, managing, and optimizing cloud resources. The No. 1 excuse that I hear from those overspending on cloud computing is that they can’t find the help they need to maximize cloud resources. They are kicking years of cloud-powered technical debt down the road, hoping that someone or some tool will come along to fix everything. ... Automation tools powered by AI can play a crucial role in ensuring that resources are only provisioned when needed and decommissioned when not in use, thus preventing idle resources from unnecessarily accruing costs. Moreover, a robust cost governance framework is essential for cloud cost management. This framework should include policies for resource provisioning, usage monitoring, and cost optimization. ... It’s frustrating that we’ve yet to learn how to do this correctly. 2020 wants their cloud spending problems back. This is not the only survey I’ve seen that reveals cost inefficacies on a massive scale. I see this myself.


The Interplay Of IoT And Critical Infrastructure Security

The vast number of interconnected devices in an IoT-driven infrastructure creates a massive attack surface. These objects often have limited processing power and may miss out on robust security features, which potentially makes them easy targets. ... The silver lining is that such heterogeneous cabling architectures can be glued together to deliver fiber-grade connectivity without the need to build new high-cost networks from scratch. An illustration of this tactic is Actelis’ hybrid-fiber technology harnessing high-performance managed Ethernet access switches and extenders to make the most of existing network infrastructures and provide gigabit speeds via virtually any wireline media. Actelis’ hybrid-fiber networking concept includes sections of fiber (for the easy-to-reach-with-fiber locations) and copper/coax that can be upgraded with Actelis’ technology to run fiber-grade communication. The company does both and provides management, security, and end-to-end integration for such entire networks, including fiber parts. This is important, as it represents a significant part of the market, selling both fiber and non-fiber networking.


MIT Researchers Introduce Generative AI for Databases

The researchers noticed that SQL didn’t provide an effective way to incorporate probabilistic AI models, but at the same time, approaches that use probabilistic models to make inferences didn’t support complex database queries. They built GenSQL to fill this gap, enabling someone to query both a dataset and a probabilistic model using a straightforward yet powerful formal programming language. A GenSQL user uploads their data and probabilistic model, which the system automatically integrates. Then, she can run queries on data that also get input from the probabilistic model running behind the scenes. This not only enables more complex queries but can also provide more accurate answers. For instance, a query in GenSQL might be something like, “How likely is it that a developer from Seattle knows the programming language Rust?” ... Incorporating a probabilistic model can capture more complex interactions. Plus, the probabilistic models GenSQL utilizes are auditable, so people can see which data the model uses for decision-making. In addition, these models provide measures of calibrated uncertainty along with each answer.


Securing Applications in an Evolving API Security Landscape

Adding further fuel to the fire, threat actors constantly innovate, developing new techniques to target APIs. This includes exploiting zero-day vulnerabilities, leveraging stolen credentials through phishing attacks, or even using bots to automate brute-force attacks against API endpoints. Traditionally, API security focused on reactive measures, patching vulnerabilities and detecting attacks after they happened. However, GenAI allows attackers to automate tasks, churning out mass phishing campaigns or crafting malicious code specifically designed to exploit API weaknesses. These attacks, known for their speed and volume, easily overwhelm traditional security solutions designed for static environments. ... APIs are crucial for modern business operations, driving innovation and customer interactions. However, without proper security measures, APIs can be vulnerable to attacks that can put sensitive data at risk, disrupt operations, and damage customer trust. ... This underscores the real-world impact of inadequate API security, including delayed innovation, frustrated customers, and lost revenue. 


Fundamentals of Descriptive Analytics

Descriptive analytics helps to describe and present data in a format that can be easily understood by a wide variety of business readers. Descriptive analytics rarely attempts to investigate or establish cause-and-effect relationships. As this form of analytics doesn’t usually probes beyond surface analysis, the validity of results is more easily implemented. Some common methods employed in descriptive analytics are observations, case studies, and surveys. ... The main disadvantage of descriptive analytics is that it only reports what has happened in the past or what is happening now without explaining the root causes behind the observed behaviors or without predicting what is about to happen in the future. The analysis is generally limited to few variables and their relationships. However, descriptive analytics becomes a powerful business resource when it is combined with other types of analytics for assessing business performance. While descriptive analytics focuses on reporting past or current events, the other types of analytics explore the root causes behind observed trends and can also predict future outcomes based on historical data analysis.


Strategies for creating seamless and personalised customer journeys

One cannot overlook the fact that CX in financial services is crucial since it influences customer satisfaction levels. In turn, this depends on providing seamless experiences across all touchpoints, beginning with the initial contact, right up to the final delivery and the consumption of financial products or services. Significantly, CX in financial services goes beyond merely offering excellent customer service. Instead, it involves ensuring frictionless, personalised experiences that surpass expectations across every stage of the customer’s journey. Be it banking, insurance or investment services, the ongoing digital transformation has substantially enhanced the expectations of customers. Today, speed, convenience and personalisation are taken for granted. As a result, financial entities that provide superior CX enjoy a clear competitive advantage in both attracting and retaining their customers, leading to increased profitability. ... Navigating complex regulatory environments ensures financial integrity. Regulations uphold standards, protecting consumers. Compliant firms earn trust, enhancing reputation alongside prioritizing customer experience for competitiveness in the market.


Cultivating Continuous Learning: Transforming L&D for the Hybrid Future and Beyond

Adapting L&D for the Digital Age requires strategic initiatives that cater to evolving demands in today's dynamic landscape. This entails embracing technology-driven learning platforms and tools that facilitate remote access, personalised learning paths, and real-time analytics. By leveraging these innovations, organisations can ensure that learning experiences are agile, responsive, and tailored to individual needs. Integrating digital collaboration tools fosters a culture of continuous learning and knowledge sharing across geographies, enhancing organisational agility and competitiveness in a digital-first world. Future-proofing L&D involves identifying and developing critical skills that will drive success in tomorrow's workplace. This proactive approach requires foresight into emerging trends and industry shifts, preparing employees with skills such as adaptability, digital literacy, creativity, and emotional intelligence. Implementing forward-thinking training programs and certifications ensures that employees remain adept and resilient in the face of technological advancements and market disruptions.


Cybersecurity Can Be a Businesses Enabler

Cybersecurity controls and protective mechanisms can protect an organization's assets - its data, people, technology equipment, etc. By actively protecting assets and preventing data breaches, an organization can avoid potential negative business impact, financial or otherwise. And because the organization does not have to worry about that potential damage, it can operate in a safe and focused fashion. ... Suffering a negative security event indicates some gap or deficiency in an organization's security posture. All organizations have security gaps, but some never report a negative security event - probably because they have invested more resources in differentiating themselves from their competitors. Implementing strong protective measures shows customers that an organization takes security seriously, which makes it a more appealing business partner. ... Customers and partners are becoming more aware of cyber risks, and they prioritize cybersecurity when they consider engaging in business. By implementing effective cybersecurity measures, a company can improve the confidence that potential customers and partners have in it. Over time, this will lead to increased loyalty and trust.


What is transformational leadership? A model for motivating innovation

The most important thing you can do as a transformational leader is to lead by example. Employees will look to you as a role model for behavior in all areas of the workplace. If you lead with authenticity, employees will pick up on that behavior and feel inspired to maintain that high standard for performance. It’s not about manipulating employees into working hard, it’s about leading by example and positively influencing others through a commitment to trust, transparency, and respect. ... To help create change, it’s important to challenge long-standing beliefs in the company and push the status quo by encouraging innovation, creativity, critical thinking, and problem-solving. Transformative leaders should help employees feel comfortable exploring new ideas and opportunities that can inject innovation into the organization. ... As a transformational leader, you will need to encourage your team to feel attached and committed to the vision of the organization. You want to ensure employees feel as committed to these goals as you do as a leader by giving employees a strong sense purpose, rather than attempting to motivate them through fear.


How Post-Quantum Cryptography Can Ensure Resilience

Quantum computing represents a major threat to data security, as it can make attacks against cryptography much more efficient. There are two ways bad actors could use this technology. One is the “Store now, decrypt later” method, in which cybercriminals steal sensitive data and wait until quantum computers have the ability to break its encryption. This is particularly important for you to know if your organization retains data with a long confidentiality span. The other method is to break the data’s digital signatures. A bad actor could “compute” credentials based on publicly available information, then impersonate someone with the authority to sign documents or approve requests. As with the above message, criminals can do this retroactively if older signatures are not updated. Today’s encryption methods cannot stand against the capabilities of tomorrow’s quantum computers. When large-scale quantum computers are built, they will have the computing ability to decrypt many of the current public key cryptography systems. 



Quote for the day:

"The only limit to our realization of tomorrow will be our doubts of today." -- Frank D Roosevelt

Daily Tech Digest - July 08, 2024

How insurtech startups are addressing the challenges of slow processes in the insurance sector

Even though compliance and regulation are critical for the security of both the insurers and customers, the regulatory process could be quite long. Compliance requirements demand meticulous attention to detail and can significantly prolong the approval process for new products and services. Another factor can be risk aversion. It (risk aversion) within the industry fosters a culture of caution, where insurers are hesitant to embrace change and experiment with new approaches to product development and underwriting. ... One of the solutions for these industrial challenges lies in the collaboration of the insurance sector and the latest technologies. Insurtech solutions offer myriad innovative tools and technologies that promise to streamline product development and automate underwriting processes. One such solution gaining traction is artificial intelligence (AI) and machine learning algorithms, which can analyse vast amounts of data in real time to assess risk and expedite underwriting decisions. 


Transforming Business Practices Through Augmented Intelligence

While AI raises apprehensions about potential job displacement, viewing it solely as a threat overlooks its capacity to enhance human capabilities, as evidenced by historical technological advancements. Training and education play a key role in this process, as AI has become an integral part of our reality and must be harnessed to its full potential. It is essential to align the use of artificial intelligence with the overall strategy of the organization for smooth integration of applications with data, processes, and collaboration between stakeholders. In a landscape where the internet simplifies transactions, software provides tools, and AI leverages data to make informed decisions, training and education become crucial. ... At its core, technology has always revolved around processing data. When viewed through the lens of enterprise architecture, an AI-powered machine learning tool can adeptly craft roadmaps tailored for businesses. Through advanced AI analytics, automation, and recommendation systems, enterprise architecture facilitates more informed and expedited decision-making processes.


Request for proposal vs. request for partner: what works best for you?

An RFProposal is an efficient choice when the nature of the work is standardized, while an RFPartner is the better choice when the buying organization is seeking a strategic partner for the overall best fit to meet its needs. ... When organizations shift to wanting to find a partner with the best possible solution, it’s important to understand the nature of the selection criteria change. With an RFPartner, buyers evaluate suppliers not only based on technical capabilities but also on the best value of the solution. ... “On the surface, an RFPartner sounds like a heavy lift, but we find that the overall time and effort is about the same,” he says. “In an RFProposal, the buyer is spending more time upfront defining the specs and in contentious negotiations. The RFPartner process flips this on its head and creates a more integrated bid solution that generates better solutions, spending more time together with the supplier co-creating, especially if your aim is making the shift to a highly collaborative vested business model to achieve strategic business outcomes.”


If you’re a CISO without D&O insurance, you may need to fight for it

D&O insurance covers the personal liabilities of corporate directors and officers in the event of incidents that lead to financial losses, reputational damage, or legal consequences. Without adequate D&O coverage, CISOs are left vulnerable, highlighting the need for this in an organization’s risk-management strategy. ... Lisa Hall, CISO at privately held Safebase, agrees that CISOs at all companies should be covered under their organizations’ D&O insurance policies, particularly in light of these new regulations. “I do think adding CISOs to D&O insurance will be more and more of a thing, and there is, for sure, more chatter in my CISO groups about how companies are handling this,” she says. “A lot of CISOs are also taking out errors and omissions insurance personally. I have that just for the consulting and advisory work I do.” ... “A lot of CISOs are thinking about this, especially after SolarWinds,” she says. “And if we feel that we’re not 100% protected for any decision we make, and we can be personally liable for a breach or possible incident even if we do the right thing, it’s really pushing CISOs to say, ‘Hey, company, I’ll join if you cover me or give me a different title.’ “


How DORA is fortifying Europe’s financial future with a new take on operational resilience

For DORA, digital operational resilience very simply means “the ability of a financial entity to build, assure, and review its operational integrity and reliability by ensuring, either directly or indirectly through the use of services provided by ICT third-party service providers, the full range of ICT-related capabilities needed to address the security of the network and information systems which a financial entity uses, and which support the continued provision of financial services and their quality, including throughout disruptions”. Developing on this statement in a conversation with FinTech Futures, Simon Treacy, a senior associate at global law firm Linklaters, describes DORA as “a very prescriptive framework for financial entities, primarily to build and improve the way that they manage ICT risk”. “It applies very broadly across the EU regulated financial sector,” he continues, “and really part of its aim is to harmonise standards so that the smallest payments firm is subject to the same rules for operational resilience as the biggest banks and insurers.”


Data Sprawl: Continuing Problem for the Enterprise or an Untapped Opportunity?

Data fabric technologies excel in integrating and managing data across various environments. However, they often focus on conventional data sources like databases, data lakes, or data warehouses. The result is a gap in integrating and extracting value from data residing in numerous SaaS applications, as they may not seamlessly fit into these traditional data repositories. The combined solution of data fabric and iPaaS can address complex business challenges, such as integrating data from SaaS applications with traditional data sources. This capability is particularly valuable in today’s business landscape, where data is increasingly scattered across various cloud and on-premises environments. The merging of data fabric and iPaaS technologies offers a groundbreaking solution to this challenge, opening the door to new opportunities in data management and analysis. The integration of data fabric with iPaaS addresses the complexity and expertise-dependency in iPaaS. Data fabric can enable users to discover, understand, and verify data before integration flows are built. 


AI’s moment of disillusionment

AI, whether generative AI, machine learning, deep learning, or you name it, was never going to be able to sustain the immense expectations we’ve foisted upon it. I suspect part of the reason we’ve let it run so far for so long is that it felt beyond our ability to understand. It was this magical thing, black-box algorithms that ingest prompts and create crazy-realistic images or text that sounds thoughtful and intelligent. And why not? The major large language models (LLMs) have all been trained on gazillions of examples of other people being thoughtful and intelligent, and tools like ChatGPT mimic back what they’ve “learned.” ... We go through this process of inflated expectations and disillusionment with pretty much every shiny new technology. Even something as settled as cloud keeps getting kicked around. My InfoWorld colleague, David Linthicum, recently ripped into cloud computing, arguing that “the anticipated productivity gains and cost savings have not materialized, for the most part.” I think he’s overstating his case, but it’s hard to fault him, given how much we (myself included) sold cloud as the solution for pretty much every IT problem.


How nation-state cyber attacks disrupt public services and undermine citizen trust

While nation-states do have advanced capabilities and visibility that are hard or impossible for cyber criminals to replicate, the general strategy for attackers is to target vulnerable perimeter devices such as VPNs or firewalls as an entry point to the network. Next they focus on obtaining privileged credentials while leveraging legitimate software to masquerade as normal activity while they scout the environments for valuable data or large repositories to disrupt. It’s important to note that the commonly exploited vulnerabilities in government IT systems are not distinctly different from the vulnerabilities exploited more broadly. Government IT systems are often extremely diverse and thus, subject to a variety of exploits. ... Currently, there are numerous policies and regulations, both domestically and internationally, which are inconsistent and vary in their requirements. These administrative requirements take significant resources which could otherwise be used to strengthen a company’s cybersecurity program. 


How Quantum Computing Will Revolutionize Cloud Analytics

As we peer into the future of quantum computing in cloud analytics, the emphasis on collaboration and continuous innovation becomes undeniable. Integrating quantum technologies with cloud systems is not just a technological upgrade but a paradigm shift requiring robust partnerships across academia, industry, and government sectors. For instance, IBM’s quantum network includes over 140 members, including start-ups, research labs, and educational institutions, working together to advance quantum computing. This collaborative model is essential because the challenges in quantum computing are not just about hardware or software alone but about creating an ecosystem that supports an entirely new kind of computing. That ecosystem comprises components such as quantum hardware development, quantum algorithms, software tools, and educational resources. Also, it has made significant achievements, such as developing quantum hardware such as the IBM Quantum System One, advancing quantum algorithms for practical applications in chemistry and materials science, and creating the Qiskit software development kit to make quantum programming more accessible.


How continuous learning is reshaping the workforce

Gone are the days when lengthy training programs were sought after and people took breaks from their careers to pick up an upskilling program. Navpreet Singh highlights that upskilling will become an ongoing process integrated into the workday. “The focus will shift from acquiring specific job skills to fostering adaptability and lifelong learning. Critical thinking, problem-solving, and creativity will be paramount as automation takes over routine tasks. Traditional ways of learning may not always reflect the skills needed. Alternative credentials, like badges and micro-credentials, will showcase the specific skills employees possess, making them more competitive. By embracing this future of upskilling, we can ensure our workforce is adaptable, future-proof, and ready to drive innovation in the ever-evolving automotive industry,” explains Singh. Within the next decade or so, we will see greater demand for agile ed-tech tools that help employees learn on the go and prepare them for new roles, says Daniele Merlerati, Chief Regional Officer APAC, Baltics, Benelux at Gi Group Holding.



Quote for the day:

"Perseverance is failing nineteen times and succeeding the twentieth." -- Julie Andrews

Daily Tech Digest - August 05, 2022

Auto Industry at Higher Risk of Cyberattacks in 2023

Connected cars are one of the most significant factors driving these risks. These vehicles feature connectivity and include autonomous features, so attackers have more potential entry points and can do additional damage once inside. Self-driving vehicle sales could reach 1 million units by 2025 and skyrocket after, so these risks will grow quickly. Automakers also face risks from connected manufacturing processes. This trend has emerged in other sectors that have embraced IT/OT convergence. One-quarter of energy companies reported weekly DDoS attacks after implementing Industry 4.0 technologies. Their attack surfaces will increase as car manufacturers likewise implement these systems. ... One of the most important changes to make is segmenting networks. All IoT devices should run on separate systems from more sensitive endpoints and data to prevent lateral movement. Encrypting IoT communications and changing default passwords is also crucial. Manufacturers should update these systems regularly, including using updated anti-malware software. 


Why Developers Need a Management Plane

A management plane empowers line developers to accomplish all of this without having a deep understanding or mastery of how to work native data plane configuration files and policies for firewalls, networking, API management and application performance management. With the management plane, platform ops teams can reduce the need for developers to build domain-specific knowledge outside the normal realm of developer expertise. For example, a management plane can have a menu of options or decision trees to determine what degree of availability and resilience an application requires, what volume of API calls can be issued against an app or service or where an app should be located in the cloud for data privacy or regulatory reasons. Equally important, the management plane can improve security by providing developers smart recommendations on good security practices or putting in place specific limits on key resources or infrastructure to ensure that developers shifting left don’t inadvertently expose their organization to serious risk.


Tech hiring enters the Big Freeze

Google and Microsoft are not the only tech companies that have started to take a more cautious approach to hiring. Earlier this year, Twitter initially issued a hiring freeze, then laid off 30% of its talent acquisition team earlier this month. At the end of June, Meta CEO Mark Zuckerberg was hostile on a call with employees, saying that “realistically, there are probably a bunch of people at the company who shouldn’t be here.” A month later, the company’s Q2 2022 financial results showed its first ever decline in revenue, with Zuckerberg telling investors that the economic climate looked even graver than it did the previous quarter. Around the same time, Apple also announced that, while the company will continue to invest in product development, it will no longer increase headcount in some departments next year. ... Research shows that employees want to be regularly offered training and the chance to develop new skills and are more likely to stay at a company if given those opportunities. The Great Resignation was a major topic of conversation in the first half of this year and, for companies that are no longer hiring, losing more employees is not an option.


Cybersecurity could offer a way for underrepresented groups to break into tech

It seems that given the sheer number of people needed in cybersecurity in the coming years could represent a way for historically underrepresented groups to find their way into tech. CJ Moses, CISO at AWS, spoke at the company keynote about the importance of diverse ways of thinking when it comes to keeping companies secure. “Another key part of our culture is having multiple people in the room with different outlooks. This could be introversion or extroversion, coming from different backgrounds or cultures, whatever enables your culture to be looking at things differently and challenging one another,” he said. He added that new ways of thinking can be transformative to cybersecurity teams. “I also think new hires can offer a team high levels of clarity because they don’t have years of bias or a group think baked into their mechanisms. So when you’re hiring, our best practices encourage being sensitive to the makeup of the interview panels, having multiple viewpoints and backgrounds, because diversity brings diversity.”


3 Things The C-Suite Should Know About Data Management And Protection

Ultimately, the massive increases in the three Vs have, by and large, resulted in inconsistent data management and protection policies in companies across the globe. So, traditional approaches to data management and protection are no longer sufficient. You need to be prepared to support empowering your IT department with the ability to meet today’s challenges. Consider solutions like autonomous data management, which uses AI-driven technology to fully automate self-provision, self-optimization and self-healing data management services for the vast amounts of data in the multi-cloud environments enterprises are migrating toward. ... The cloud makes a lot of sense for a lot of reasons. It’s flexible, with scalability and mobility; efficient, including its accessibility and speed to market; and cost-effective, as it includes pay-as-you-go models and helps eliminate hardware expenses. But it can be a fickle beast, especially in this ever-increasingly multi-cloud world. This refers to how enterprise data is being dispersed across on-premises data centers and the many private and public cloud service providers.


5 best practices for secure collaboration

What we have seen is that has rapidly changed now over the last couple of years as calling is still obviously very important, but other collaboration technologies have entered the landscape and have become equally, if not arguably, more important. And the first one of those is video. The challenges, when you think about securing video, obviously a lot of folks have heard about unauthorized people [discovering] a meeting and [joining] it with an eye toward potentially disrupting the meeting or toward snooping on the meeting and listening in. And that has, fortunately, been addressed by most of the vendors. But the other real concern that we have seen arise from a security and especially a compliance perspective is meetings are generating a lot of content. ... If you are a CSO, obviously you have ultimate responsibility for collaboration security. But you also want to work with the collaboration teams to either delegate ownership of managing day-to-day security operations to those folks or working with them to get input into what the risks are and what are the possible mitigation techniques. 


Build .NET apps for the metaverse with StereoKit

Developing with StereoKit shouldn’t be too hard for anyone who’s built .NET UI code. It’s probably best to work with Visual Studio, though there’s no reason you can’t use any other .NET development environment that supports NuGet. Visual Studio users will need to ensure that they’ve enabled desktop .NET development for Windows OpenXR apps, UWP for apps targeting HoloLens, and mobile .NET development for Oculus and other Android-based hardware. You’ll need an OpenXR runtime to test code against, with the option of using a desktop simulator if you don’t have a headset. One advantage of working with Visual Studio is that the StereoKit development team has provided a set of Visual Studio templates that can speed up getting started by loading prerequisites and filling out some boilerplate code. Most developers are likely to want the .NET Core template, as this works with modern .NET implementations on Windows and Linux and gets you ready for the cross-platform template under development. 


The Data science journey of Amit Kumar, senior enterprise architect-deep learning at NVIDIA

The most important thing for aspirants is to get the fundamentals right before diving into data science and AI. Having a basic but intuitive understanding of linear algebra, calculus, and information theory helps to get a faster grip. Aspiring data scientists should not ignore fundamental principles of software engineering, in general, because nowadays the market is looking for full-stack data scientists with the capability to build an end-to-end pipeline, rather than just being a data science algorithm expert. ... My biggest challenge, which ultimately turned into my biggest achievement, was to start from scratch and build a world-class center of excellence in data science at HP India along with Niranjan Damera Venkata, Madhusoodhana Rao and Shameed Sait. This challenge was turned into an achievement by going into the start-up mode within HP. Though we were part of a large organisation, we made sure that the center of excellence operates the way a successful startup works by inculcating the culture of mutual respect and healthy competition, attracting and hiring best talents, and providing freedom and flexibility.


Confidential Computing with WebAssembly

Confidential computing is of particular use to organizations that deal in sensitive, high value data — such as financial institutions, but also a wide variety of organizations. “We felt that confidential computing was going to be a very big thing be that it should be easy to use,” said Bursell, was then chief security architect in the office of Red Hat’s chief technology officer. “And rather than having to rewrite all the applications and learn how to use confidential computing, it should be simple.” But it wasn’t simple. Among the biggest puzzles: attestation, the mechanism by which a host measures a workload cryptographically and communicates that measurement to a third party. “One of the significant challenges that we have is that all the attestation processes are different,” said McCallum, who led Red Hat’s confidential computing strategy as a virtualization security architect. “And all of the technologies within confidential computing are different. And so they’re all going to produce different cryptographic caches, even if it’s the same underlying code that’s running on all.”


The Computer Scientist Challenging AI to Learn Better

The most successful method, called replay, stores past experiences and then replays them during training with new examples, so they are not lost. It’s inspired by memory consolidation in our brain, where during sleep the high-level encodings of the day’s activities are “replayed” as the neurons reactivate. In other words, for the algorithms, new learning can’t completely eradicate past learning since we are mixing in stored past experiences. There are three styles for doing this. The most common style is “veridical replay,” where researchers store a subset of the raw inputs — for example, the original images for an object recognition task — and then mix those stored images from the past in with new images to be learned. The second approach replays compressed representations of the images. A third far less common method is “generative replay.” Here, an artificial neural network actually generates a synthetic version of a past experience and then mixes that synthetic example with new examples. My lab has focused on the latter two methods. 



Quote for the day:

"The leadership team is the most important asset of the company and can be its worst liability" -- Med Jones