Resource-Aware DYnamically Adaptable machine Learning – RADYAL

Resource-Aware DYnamically Adaptable machine Learning

In this project we propose an original interdisciplinary approach that allows DNN models to be dynamically configurable at run-time on a given reconfigurable hardware accelerator architecture, depending on the external environment, following an approach based on feedback loops and control theory.

Run-time configurable and contrôlable DNN SW and HW accel

At software (SW) level, for a given DNN model, different variants with incremental precision levels can be obtained by setting parameters along different dimensions: (i) data precision or quantization (increasing/decreasing bit-width of activations and/or weights), (ii) degree of sparsification (e.g., pruning, tensor decomposition), (iii) depth of the NN (number and type of network layers to execute). Depending on the chosen SW precision level, the mean output accuracy will change, as well as energy consumption and timing.<br />The key observation is that, for some particularly “easy” inputs, using high-precision energy-hungry computations is an “overkill”. Conversely, for “hard” inputs, low-precision energy-efficient computations are not enough. Therefore, being able to dynamically change the SW precision is key to enable energy-efficient and accurate NN computations. At the same time, at the hardware (HW) level, the DNN accelerator needs to be configurable at runtime to satisfy SW processing requirements. Different degrees of existing HW strategies have an impact on energy consumption and runtime. For example, changing operation mapping strategies (e.g., Input-Stationary, Weight-Stationary, Output-Stationary, Row-Stationary), the number of active Processing Elements (PE) inside an accelerator (e.g., 16x16 or 8x8 systolic array), their data precision (e.g., 8 bits, 4 bits), sparsification support (e.g., weight pruning, element-wise pruning, structured pruning, dynamic pruning, activation sparsification).<br />We want to explore the different configuration opportunities at SW and HW level – independently and jointly – to be able to build a light surrogate model of the configurations’ impact on system’s attributes. This model will enable a control-theory-oriented approach to combine SW and HW configurations delivering accurate DNN outputs while ensuring the most efficient computation possible, given the runtime conditions.