Question: How often do random noises sound like farts?
Short answer: 5.5% of the time.
Long answer: In previous posts, we have documented the surprising fact that nearly all fart sounds share similar sound frequencies. This is the basis of our Flatus Reflector technology, which detects farts in arbitrary audio samples.
We have used this technology to search for lost farts in audio archives and to exonerate those accused unfairly of farting in public. But in general, the application of this technology has been limited by the fact that the Reflector is very computationally intensive. Even using the College’s the supercomputing cluster, we have been able to examine only a miniscule fraction of the audio available to us.
We have therefore tried a different approach to fart detection, one that relies on the power of artificial neural networks (ANNs). Some time ago, we showed that ANNs could be used to identify individuals accurately from the sound of their farts. Fart research with ANNs goes back to at least 1991.
Using our extensive fart database, we trained a three-layer neural network to discriminate between fart and non-fart sounds. As in previous work, we reduced the dimensionality of the input by expressing each sound input in terms of its Mel-frequency cepstrum. This yields a 13-dimensional input for each sound, which is then processed by a hidden layer, which provides output to a single node that outputs the probability that the sound was a fart:
The network was trained on 230 farts and 230 non-farts, the latter being comprised of freely available clips found on the internet. The network was trained with gradient descent and tested on a cross-validation dataset.
Performance typically ranged between 75.8% and 91.4% accuracy, with little difference being observed as the number of hidden units varied between 10 and 100. Interestingly, there were far more false negatives (14.6%) than false positives (5.5%).
Among the false positives, the most significant involved someone dropping a piece of metal, which was classified as having an 87% chance of being a fart
Strangely, the network classified a rooster crowing as 82% likely to be a fart:
Among the false negatives were a number of farts for which the audio recording was distorted in some way, such as:
But other false negatives were clearly farts, such as the following, which was assigned an 18.7% chance of being a fart:
Overall, these results are consistent with the well-known fact that ANNs often perceive sensory stimuli in ways that depart significantly from those of biological neural networks. Nevertheless, our first attempt at an AI based solution to fart detection was largely successful. Because ANNs, once trained, are relatively simple to deploy, our FartNet algorithm should soon become our standard approach to fart detection.