It is the Embedded Vision Summit. Every year this event gets bigger, reflecting the growing interest in the area. Silicon is now capable enough that it is feasible to do complex algorithms in smartphones and automotive processors, rather than requiring an upload to the cloud. Almost overnight, machine learning (sometimes called deep learning) has become a hot topic. In fact, in 2014 machine learning was not even on the Gartner Hype Cycle for Emerging Technology and by 2015 it had climbed all the way to Peak Hype. Hopefully, next year we not be in the Trough of Dissillusionment. Yesterday, Chris Rowen, the CTO of Cadence's IP group, gave a talk at the Embedded Vision Summit titled "The Road Ahead for Neural Networks: Five Likely Surprises." Actually, like Monty Python's Spanish inquisition, he had six. Machine Learning technologies are driving new ecosystems for deep learning. Depending on what is included and how optimistic you are, the values for the total size of the market go from $5B to $2T. A complete value chain for automotive image recognition, for example, can involve silicon IP suppliers such as Cadence, SoC designers, foundries, owners of training data, system integrators (often called Tier 1s in automotive terminology), automotive manufacturers (aka OEMs), and the end-user out of who, at least indirectly, those $ have to originate. Today, most machine learning is about recognition, but there is starting to be an increasing about of work on training for judgement and strategy. Although it is not a practical problem that most of us have to deal with every day, it is significant that Google AlphaGo won 4-1 against the best human Go player using two neural networks (one to find good moves, one to evaluate board positions). One of the challenges with this sort of problem is getting enough good data. Since Go games have been recorded for years (centuries?), there is a lot of data available but most areas do not have an existing body of data like that. I said at the beginning that interest in this area is exploding since silicon has become capable enough to handle real problems, but Chris pointed out in his talk that neural networks are still too inefficient: they need too much memory bandwidth, too much compute, and are not accurate enough, not to mention they are too energy thirsty. Good neural networks need lots of compute (especially DSP multiply-add). Cadence Tensilica Vision DSPs are especially good for this, giving greater efficiency than either GPUs or FPGAs. During his talk, Chris introduced the latest Vision P6 DSP. You can read a lot more about it in my blog posted yesterday . Before he got to the surprises, Chris had six non-surprises: Neural networks will continue to proliferate in cloud-based applications (Google, Facebook, Baidu) Neural networks will expand rapidly into real-time embedded functions (ADAS, smarterphones) Power constraints and extreme throughput needs will drive CNN optimization in processor platforms, both embedded and server Real-time neural networks will evolve from object recognition to action recognition (autonomous vehicles) Vision-based problems are/will dominate the computations and the high-profile deployments Expect a mad, sometimes unguided, scramble for expertise, data, and applications. So what were the five/six surprises? >100X energy and >20X in bandwidth from neural network and engine architecture near-term In time: deployment of 1000 tera-MAC (peta-MAC) embedded and 1,000,000 tera-MAC (exa-MAC) server neural networks Network optimization evolves from ad hoc exploration to automated "synthesis", a new kind of EDA New value changes will emerge, and swing between vertical integration and disintegration Data is king, so ccess to large, diverse training datasets will create new winners Potential backlash over privacy and "rise of the machines" (Skynet?) Lest anyone gets too excited, here is a cautionary note from Yann LeCun, Facebook's AI chief (and a keynoter at the Embedded Vision Summit a couple of years ago) about AlphaGo's victory: As I've said in previous statements: most of human and animal learning is unsupervised learning. If intelligence was a cake, unsupervised learning would be the cake, supervised learning would be the icing on the cake, and reinforcement learning would be the cherry on the cake. We know how to make the icing and the cherry, but we don't know how to make the cake. There's a long way to go before our smartphones are really smart, but the technology is improving every year. Previous: New Algorithms for Vision Require a New Processor Next: DAC: One Month and Counting![]()
![]()
↧