To run that tiny neural network, a microcontroller also needs a lean inference engine. A typical inference engine carries some dead weight — instructions for tasks it may rarely run. The extra code poses no problem for a laptop or smartphone, but it could easily overwhelm a microcontroller. “It doesn’t have off-chip memory, and it doesn’t have a disk,” says Han. “Everything put together is just one megabyte of flash, so we have to really carefully manage such a small resource.” Cue TinyEngine. The researchers developed their inference engine in conjunction with TinyNAS. TinyEngine generates the essential code necessary to run TinyNAS’ customized neural network. Any deadweight code is discarded, which cuts down on compile-time. “We keep only what we need,” says Han. “And since we designed the neural network, we know exactly what we need. That’s the advantage of system-algorithm codesign.” In the group’s tests of TinyEngine, the size of the compiled binary code was between 1.9 and five times smaller than comparable microcontroller inference engines from Google and ARM. TinyEngine also contains innovations that reduce runtime, including in-place depth-wise convolution, which cuts peak memory usage nearly in half. After codesigning TinyNAS and TinyEngine, Han’s team put MCUNet to the test.
The breadth of database systems available today is staggering. Something like Cassandra lets us store a huge amount of data for the amount of memory the database is allocated; Elasticsearch is different, providing a rich, interactive query model; Neo4j lets us query the relationship between entities, not just the entities themselves; things like Oracle or PostgreSQL are workhorse databases that can morph to different types of use case. Each of these platforms has slightly different capabilities that make it more appropriate to a certain use case but at a high level, they’re all similar. In all cases, we ask a question and wait for an answer. This hints at an important assumption all databases make: data is passive. It sits there in the database, waiting for us to do something. This makes a lot of sense: the database, as a piece of software, is a tool designed to help us humans — whether it's you or me, a credit officer, or whoever — interact with data. But if there's no user interface waiting, if there's no one clicking buttons and expecting things to happen, does it have to be synchronous? In a world where software is increasingly talking to other software, the answer is: probably not.
Data lakes in the cloud have high durability, low cost, and unbounded scale, and they provide good support for the data science and machine learning use cases that many enterprises prioritize today. But, all the traditional analytics use cases still exist. Therefore, customers generally have, and pay for, two copies of their data, and they spend a lot of time engineering processes to keep them in sync. This has a knock-on effect of slowing down decision making, because analysts and line-of-business teams only have access to data that’s been sent to the data warehouse rather than the freshest, most complete data in the data lake. ... The complexity from intertwined data lakes and data warehouses is not desirable, and our customers have told us that they want to be able to consolidate and simplify their data architecture. Advanced analytics and machine learning on unstructured and large-scale data are one of the most strategic priorities for enterprises today, – and the growth of unstructured data is going to increase exponentially – therefore it makes sense for customers to think about positioning their data lake as the center of data infrastructure. However, for this to be achievable, the data lake needs a way to adopt the strengths of data warehouses.
Apache Airflow, an open source workflow management tool, is rapidly being adopted by many businesses for the management of ETL processes and machine learning pipelines. Many large tech companies such as Google and Slack are using it and Google even built their cloud composer tool on top of this project. I am noticing Airflow being mentioned more and more often as a desirable skill for data scientists on job adverts. As mentioned at the beginning of this article I believe it will become more important for data scientists to be able to build and manage their own data pipelines for analytics and machine learning. The growing popularity of Airflow is likely to continue at least in the short term, and as an open source tool, is definitely something that every budding data scientist should at learn. ... Data science code is traditionally messy, not always well tested and lacking in adherence to styling conventions. This is fine for initial data exploration and quick analysis but when it comes to putting machine learning models into production then a data scientist will need to have a good understanding of software engineering principles. If you are planning to work as a data scientist it is likely that you will either be putting models into production yourself or at least be involved heavily in the process.
The payment instruction itself is a message to the partner bank, which then triggers a normal UPI transaction from the customer’s designated UPI bank to the destination partner bank through the National Payments Corporation of India (NPCI). The destination partner bank forwards the payment to the addressee’s default UPI bank registered with WhatsApp. A confirmation of credit is also sent through WhatsApp and reaches the message box of the recipient. It is possible that at either end, the WhatsApp partner bank may not be the customer’s bank. Hence, there may be the involvement of four banks, the NPCI and WhatsApp in completing the transaction. As far as the user is concerned, the system is managed by WhatsApp and none of the other players is visible. Though WhatsApp is not licensed to undertake UPI transactions directly, it engages the services of its partner banks to initiate the transaction. As these partner banks are not bankers for the customers, they engage two more banks to assist them. Finally, NPCI acts as the agent of the two banks through which the money actually passes through to the right bank. Thus, there is a chain of principal agent transaction and the roles of the customer, WhatsApp, banks, etc., need to be clarified.
New Circuit Compression Technique Could Deliver Real-World Quantum Computers Years Ahead of Schedule
“By compressing quantum circuits, we could reduce the size of the quantum computer and its runtime, which in turn lessens the requirement for error protection,” said Michael Hanks, a researcher at NII and one of the authors of a paper, published on November 11, 2020, in Physical Review X. Large-scale quantum computer architectures depend on an error correction code to function properly, the most commonly used of which is surface code and its variants. The researchers focused on the circuit compression of one of these variants: the 3D-topological code. This code behaves particularly well for distributed quantum computer approaches and has wide applicability to different varieties of hardware. In the 3D-topological code, quantum circuits look like interlacing tubes or pipes, and are commonly called “braided circuits. The 3D diagrams of braided circuits can be manipulated to compress and thus reduce the volume they occupy. Until now, the challenge has been that such “pipe manipulation” is performed in an ad-hoc fashion. Moreover, there have only been partial rules for how to do this. “Previous compression approaches cannot guarantee whether the resulting quantum circuit is correct,” said co-author Marta Estarellas, a researcher at NII.
“Remember that all your attacker cares about is stealing passwords...That’s a key difference between hypothetical and practical security.” — Microsoft’s Alex Weinert In other words, the bad guys will do whatever is necessary to steal your password and a strong password isn’t an obstacle when criminals have a lot of time and a lot of tools at their disposal. ... MFA based on phones, aka publicly switched telephone networks or PSTN, is not secure, according to Weinert. (What is typical MFA? It’s when, for example, a bank sends you a verification code via a text message.) “I believe they’re the least secure of the MFA methods available today,” Weinert wrote in a blog. “When SMS (texting) and voice protocols were developed, they were designed without encryption...What this means is that signals can be intercepted by anyone who can get access to the switching network or within the radio range of a device,” Weinert wrote. Solution: use app-based authentication. For example, Microsoft Authenticator or Google Authenticator. An app is safer because it doesn’t rely on your carrier. The codes are in the app itself and expire quickly.
Once the immediacy of both the pandemic dissipates and the political heat cools, cybersecurity issues will likely surface again in new or revived legislation in many states, even if weaved throughout other related matters. It’s difficult to separate cybersecurity per se from adjoining issues such as data privacy, which has generally been the biggest topic to involve cybersecurity issues at the state level over the past four years. “You really don’t have this plethora of state cybersecurity laws that would be independent of their privacy law brethren,” Tantleff said. According to the National Conference of State Legislatures, at least 38 states, along with Washington, DC, and Puerto Rico introduced or considered more than 280 bills or resolutions that deal significantly with cybersecurity as of September 2020. Setting aside privacy and some grid security funding issues, there are two categories of cybersecurity legislative issues at the state level to watch during 2021. The first and most important is spelling out more clearly what organizations need to meet security and privacy regulations. The second is whether states will pick up election security legislation left over from the 2020 sessions.
Human error is the main cause of security breaches, wrong data interpretation, mistaken insights, and a variety of other damning experiences the insights industry has had to wade through ever since its conception. Zooming out to take a wider look, human error is the cause of mistaken elections, aviation accidents, cybersecurity issues, etc. but also scientific breakthroughs across the world. While some mistakes yield true results, most have dangerous consequences that could have been avoided if we were more careful. To err is human, but in an industry where mistakes have real-world consequences, to err is to potentially cost a business it’s life. If we stick with the artificial intelligence and automation example, automated processes with next generation technology are the most poignant example of humans trying to make up for their mistakes and can help minimise human error at all stages ... The main benefit of combining human oversight with this next generation technology, is that we can catch and fix any bugs that arise before they harm the research process and projects that rely on said technology. But we need to be wary that humans cannot catch every mistake, and when one slips through that is when oversight takes on a whole new, disappointing meaning.
The decision on where to train and deploy AI models can be determined by balancing considerations across six vectors: scalability, latency, autonomy, bandwidth, security, and privacy. In terms of scalability, in a perfect world, we’d just run all AI workloads in the cloud where compute is centralized and readily scalable. However, the benefits of centralization must be balanced out with the remaining factors that tend to drive decentralization. For example, if you depend on edge AI for latency-critical use cases and for which autonomy is a must, you would never make a decision to deploy a vehicle’s airbag from the cloud when milliseconds matter, regardless of how fast and reliable your broadband network may be under normal circumstances. As a general rule, latency-critical applications will leverage edge AI close to the process, running at the Smart and Constrained Device Edges as defined in the paper. Meanwhile, latency-sensitive applications will often take advantage of higher tiers at the Service Provider Edge and in the cloud because of the scale factor. In terms of bandwidth consumption, the deployment location of AI solutions spanning the User and Service Provider Edges will be based on a balance of the cost of bandwidth, the capabilities of devices involved and the benefits of centralization for scalability.
Quote for the day:
"If you want to do a few small things right, do them yourself. If you want to do great things and make a big impact, learn to delegate." -- John C. Maxwell