So today I'm going to present and share my research on efficient and robust hardware

for neural networks with you.

So neural networks were proposed originally to simulate the synapse network in our human

brain.

For example, this figure shows a very simple neural network.

The computations at the neural includes the multiplication operations with the input and

weight and also the cumulation results will be added with the BIOS and processed by a

nonlinear activation function.

But with the increasing depth of neural networks, massive multiplication and addition operations,

or we call it micro-operations, are required to be executed in neural networks.

For example, in the previous year, GPT-3, a large-amplified model neural network, was

used in child GPT, consisting of 96 layers with 175 billion weights.

So running this large-amplified model triggers the training of micro-operations.

And training GPT, of course, consumes a lot of energy consumption.

And which is also comparable to the electricity consumption of 120 years for average US households.

And inference, where pre-treat neural networks were used to do predictions, consumes even

more energy consumption.

According to Alidia and Amazon, 80% of this energy is used for inference.

So my group is trying to develop some methodologies from AI efficient AI algorithms and efficient

hardware to reduce energy consumption.

So in this talk, I'm going to present two efficient AI algorithms, class aware pruning

and class exclusion early exit.

And also I will present some efficient methodologies to reduce the digital accelerators for neural

networks like the systolic array in TPUs.

And the energy consumption in accelerating neural networks, majorly from the data movement

by loading in the waste, and also the huge number of micro-operations.

So that to reduce the data movement, I further developed a logic-based neural network design.

So after cleaning a neural network, we embed the waste into the circuits and optimize the

circuits so that the circuits can be used for application-specific design.

And we also target some in-memory computing platforms like R1-based course bars.

So they are in-memory computing platforms.

They don't need to load the waste from the memories.

They just do the computing and save the data in the memory.

But they are not mature enough.

They suffer from variations.

So if you do not address the variations, the accuracy of the neural networks accelerated

by the course bars will be degraded significantly.

So I will present a topic on this.

So let's first look at the first topic about pruning.

So you might have heard of a lot of pruning techniques.

But our work proposed a new pruning perspective called class-aware pruning.

So the basic concept can be seen from this figure.

The basic idea is that if you use a very simple neural network to do three classifications,

and we found that different neurons, they are important for different number of classes.

So for example, this neuron is important for three classes.

So important means that if you remove this neuron, the accuracy of those three classes

will be reduced significantly.

And this neuron is important for two classes, for example.

And this neuron is important for one class.

And this neuron is important for non-classes.