Page 314 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

6. Discussion and Future Work 307

Table 15.3 The Evolved Network Improves Over the Hand-Designed
Baseline When Trained on MSCOCO Alone
Model BLEU-4 CIDEr METEOR
DNGO [34] 26.7 d d
Baseline [49] 27.7 85.5 23.7
Evolved 29.1 88.0 23.8

FIGURE 15.6
Results for captions generated by an evolved model for the online magazine images rated
from 1 to 4, with 4 ¼ Correct, 3 ¼ Mostly Correct, 2 ¼ Mostly Incorrect, 1 ¼ Incorrect.
Left: On iconic images, the model is able to get about one half correct; Right: On all
images, the model gets about one fifth correct. The superior performance on iconic
images shows that it is useful to build supplementary training sets for specific image
types.

6. DISCUSSION AND FUTURE WORK
The results in this paper show that the evolutionary approach to optimizing deep
neural networks is feasible: The results are comparable to hand-designed architec-
tures in benchmark tasks, and it is possible to build real-world applications based
on the approach. It is important to note that the approach has not yet been pushed
to its full potential. It takes a couple of days to train each deep neural network on
a state-of-the-art GPU, and over the course of evolution, thousands of them need
to be trained. Therefore, the results are limited by the available computational
power.
Interestingly, since it was necessary to train networks only partially during evo-
lution, evolution is biased towards discovering fast learners instead of top per-
formers. This is an interesting result on its own: evolution can be guided with
goals other than simply accuracy, including training time, execution time, or mem-
ory requirements of the network. On the other hand, if it was possible to train the