0
\$\begingroup\$

So i've been trying to find out the answer to this question but i can't seem to find a specific answer with examples. by the way i have heard about "Machine Learning Processors" but i don't know what is the difference between them and ordinary Processors.

So overall i have two questions:

1- In what areas of Electrical Engineering is Machine Learning/Artificial Intelligence applied?

2- What are "Machine Learning Processors" or "AI Chips"?

\$\endgroup\$
0

2 Answers 2

4
\$\begingroup\$

First, let's get some terminology straight.

  • AI, Artificial Intelligence, is using algorithms that mimic human intelligence
  • ML, Machine Learning, is using feedback to improve (train) AI algorithms
  • DL, Deep Learning, is using self-generated feedback for training

DL is a subset of ML, is a subset of AI.

For (1), I think we need to draw a distinction between using AI to do engineering work, and using it to implement systems.

  • AI techniques are making inroads into Electronic Design Automation, or EDA. More about that here: https://www.electronicdesign.com/industrial-automation/article/21120058/eda-in-the-era-of-ai

  • Another hot topic for AI in electronics is antenna design. Lots of research and practical applications in this area.

  • AI systems applications on the other hand are seeing wide use for automation, signal processing, big data inference extraction and numerous other areas - way too many to list here.

For (2), AI processors are architectures optimized to run AI algorithms, commonly referred to as neural networks. Generally, they support a large number of parallel computations and may also have specific hardware for implementing core AI-specific neural-net calculations, such as non-linear thresholding (e.g., ReLU) that models how neurons process inputs.

AI processor chip examples include DSP, GPU, FPGA, and AI-specific ASICs. Each has their own trade-off in flexibility in the range of AI problem classes they can address vs. power and cost.

A server used for network development / network training wants flexibility (efficient feedback for training, fast feed-forward for inference), while an embedded application like autonomous driving can use a more restrictive architecture optimized for a specific network type and feed-forward inference only.

\$\endgroup\$
3
  • \$\begingroup\$ I would not put too much weight on the accuracy of definitions given here but that does not mean it is not otherwise a good answer. \$\endgroup\$ May 10, 2020 at 4:22
  • \$\begingroup\$ I felt it important to clarify at least AI vs. ML as the OP was using those terms but likely not understanding how they relate to each other, or what ‘machine learning’ actually entails (training on datasets.) Going further than that basic definition... then the Q should migrate to the AI SE. \$\endgroup\$ May 10, 2020 at 6:08
  • \$\begingroup\$ I agree with that \$\endgroup\$ May 10, 2020 at 17:05
1
\$\begingroup\$

Freiend of mine used neural-system-waveform-processing, to classify up to 8 different resonant modes of an impulse-response system with moderate "Q" physical modes. Using only ONE Centrally mounted sensor.

The waveforms differed because of how energy reflected off mechanical boundaries.

Key was a lot of patience, many hundreds of "learning/testing" cycles, and the determination of

  • HOW DEEP THE # of LAYERS,

  • HOW WIDE TO MAKE EACH LAYER

  • WHAT NONLINEARITY TO USE IN EACH LAYER

  • HOW CLEAN MUST THE SENSOR BE (what SNR)

  • HOW TO IMPLEMENT INTERFERENCE CANCELLATION to better than 12 bits

  • WHAT DENSITY OF TIME SAMPLING WAS REQUIRED

  • WHAT FRACTION OF THE WAVEFORM TO INCLUDE (onset? and decay?)

As result, with multi-threading the recognition code, a TENSY was adequate to handle 10,000 samples/second (100 uS sample spacing) for 100 adjacent samples.

In realtime, with acceptable latency.

Using datasets with thousands of waveforms (quite an effort to gather and classify those), with 60% for training and 40% for testing, took several hours on a 4-processor Pentium to reach 95% accuracy.

The skill acquisition was

  • 3 months on genetic algorithms

  • 3 months with online Artificial Neural Systems Theory videos

  • 6 months evolving the thresholding/nonlinearity, depth, width, #samples, the pursuit of C++ parallelism, and methods to visually observe (computer oscilloscope) waveforms and convergences of learning.

\$\endgroup\$

Not the answer you're looking for? Browse other questions tagged or ask your own question.