Activation functions for deep neural networks
Hardware for implementing a Deep Neural Network (DNN) for performing an activation function includes, at an activation module for performing an activation function, a programmable lookup table for storing lookup data approximating the activation function over a first range of input values to the activation module, the method comprising: providing calibration data to a representation of the hardware; monitoring an input to an activation module of the representation of the hardware so as to determine a range of input values to the activation module; generating lookup data for the lookup table representing the activation function over the determined range of input values; and loading the generated lookup data into the lookup table of the hardware, thereby configuring the activation module of the hardware for performing the activation function over the determined range of input values.
1. Hardware for implementing a Deep Neural Network (DNN) comprising an activation module for performing an activation function and a configuration module for determining a first range of input values to the activation module, the configuration module being configured to:
cause calibration data to be provided to the hardware wherein the hardware is configured to implement the DNN on the calibration data; and
monitor an input to the activation module of the hardware during implementation of the DNN on the calibration data so as to determine the first range of input values to the activation module; the activation module comprising:
a programmable lookup table for storing lookup data representing an activation function over the first range of input values to the activation module; and
a Rectified Linear Unit (ReLU) operable to perform a ReLU activation function;
wherein the configuration module is further configured to, on the DNN requiring the activation module to implement a ReLU activation function, cause the activation module to perform the activation function at the ReLU and turn off the programmable lookup table.
2. Hardware as claimed in claim 1 , wherein the configuration module is configured to, on the DNN requiring the activation module to implement an activation function using the programmable lookup table, cause the activation module to perform the activation function using the programmable lookup table and turn off the ReLU.
3. Hardware as claimed in claim 1 , wherein, on the DNN requiring the activation module to implement an activation function using the programmable lookup table, the ReLU is configured to clamp input values received at the activation module which lie outside the first range of input values, the clamped input values being subsequently passed to the programmable lookup table for implementation of the activation function.
4. Hardware as claimed in claim 3 , wherein the ReLU is configured to clamp the input values at the closest extreme of the first range of input values.
5. Hardware as claimed in claim 1 , wherein the configuration module is configured to form a histogram of input values representing the probability of occurrence of input values and use as the bounds of the first range of input values a pair of input values between which a predefined or programmable proportion of the distribution of input values lies.
6. Hardware as claimed in claim 1 , wherein the hardware is embodied on an integrated circuit.
7. Hardware as claimed in claim 1 , wherein the configuration module is further configured to generate lookup data for the programmable lookup table representing the activation function over the first range of input values.
8. A method for implementing in hardware a Deep Neural Network (DNN), the hardware comprising an activation module for performing an activation function and a configuration module for determining a first range of input values to the activation module, the activation module comprising, a programmable lookup table for storing lookup data representing an activation function over a first range of input values to the activation module, and a Rectified Linear Unit (ReLU) operable to perform a ReLU activation function, the method comprising:
causing calibration data to be provided to the hardware;
monitoring an input to the activation module of the hardware during implementation of the DNN on the calibration data so as to determine the first range of input values to the activation module; and
when the DNN requires the activation module to implement a ReLU activation function:
turning off the programmable lookup table; and
causing the activation module to perform the activation function at the ReLU.
9. The method as claimed in claim 8 , further comprising:
when the DNN requires the activation module to implement an activation function using the programmable lookup table:
clamping the input values received at the activation module which lie outside of input values at the closest extreme of the first range of input values;
subsequently passing the clamped input values to the programmable lookup table; and
causing the activation module to perform the activation function using the programmable lookup table.
10. The method of claim 8 , wherein the method further comprises:
generating lookup data for the programmable lookup table representing the activation function over the first range of input values.
11. A non-transitory computer readable storage medium having stored thereon a computer readable dataset description of hardware for implementing a Deep Neural Network (DNN) comprising an activation module for performing an activation function and a configuration module for determining a first range of input values to the activation module, that, when processed in an integrated circuit manufacturing system, causes the integrated circuit manufacturing system to manufacture an integrated circuit embodying the hardware, wherein the configuration module is configured to cause calibration data to be provided to the hardware wherein the hardware is configured to implement the DNN on the calibration data and monitor an input to the activation module of the hardware during implementation of the DNN on the calibration data so as to determine the first range of input values to the activation module, the activation module comprising:
a programmable lookup table for storing lookup data representing an activation function over the first range of input values to the activation module; and
a Rectified Linear Unit (ReLU) operable to perform a ReLU activation function;
wherein the configuration module is configured to, on the DNN requiring the activation module to implement a ReLU activation function, cause the activation module to perform the activation function at the ReLU and turn off the programmable lookup table.