Words: Emiko Wijeysundera (she/her)
There is a common phrase in computer science: ‘garbage in, garbage out’. It means that the quality of a program’s outputs is only as good as the inputs entered. The concept is especially relevant for Artificial Intelligence (AI) which gains its “intelligence” from analysing large amounts of data. If the data is invalid, misrepresentative, or untrue, the learning is baseless. That “intelligence” could even be harmful. The artificial appetite is hungry for data, and garbage won’t do.
Let’s break this down. What exactly is Artificial Intelligence? According to ChatGPT, (because who better to ask?), ‘Artificial intelligence (AI) refers to the development of computer systems or machines that can perform tasks typically requiring human intelligence. These tasks include learning, reasoning, problem-solving, understanding language, recognizing patterns, and making decisions’. Computer systems can achieve these behaviours through virtual architectures called neural networks (NN) designed to mimic the human brain.
Neural networks are not magic; they are maths. Like an equation, they take in a numerical input and produce a numerical output. Inputted data is represented by a number, which is passed through a series of interconnected “neurons”. Each neuron manipulates the input with different parameters, called weights, and biases (think mathematical operations like multiplication and addition), before producing an output in the final layer. The NN improves through a process called Machine Learning. This operates in a similar way to traditional school examinations, where academic material is tested through exams then corrected. The network is fed input data (asked questions), and the produced outputs (written answers) are compared to the correct outputs (an answer key). Based on the NN’s performance (or score, in our exam analogy), the network’s weights and biases are adjusted. Over many tests the performance improves, and the parameters settle to values that produce correct outputs. The mathematical equation that is our NN now produces the “correct answer”.
But there is a catch. In school, if you specialise in organic chemistry you can’t be expected to ace an exam on Shakespeare. Our network will produce the correct answer for the data it was trained on. Perhaps this seems obvious; if our network is trained to classify dogs and cats, it cannot be expected to differentiate horses and fish. The point is even more particular – if we want our network to classify dogs and cats, we must train it on a comprehensive and diverse set of dogs and cats. A network trained only on cats and Golden Retrievers will be able to differentiate cats from, you guessed it, Golden Retrievers, but not cats from Bulldogs or Dalmatians. For any other type of dog the NN is useless – garbage out.
The well-trained network has a huge appetite for a robust and diverse dataset that covers all cases. The appetite appears insatiable (and this may be correct), but in the digital age data is everywhere. Cookies follow your browsing history, every social media like is documented, GPS signals are tracked; technology produces hordes of data. The key is good quality, representative, and accurate data. As we have learned, networks trained on false or misleading data are not accurate. In fact, they can be quite dangerous. Machine Learning operates on pattern recognition, with each parameter mapping a shared characteristic between data. If we are not careful, this can be harmful. The data we produce and collect often reflects society’s stereotypes. When NNs train on biassed data, dangerous generalisations can be perpetuated and exacerbated. This is especially problematic when AI is trusted or relied on to be accurate and factual.
One poignant example lies in the practice of preventative policing, which uses AI to predict the locations of future crimes and make predictions about offenders in the criminal justice system. The justice system has a long history of systematic racism which is evident in arrest, sentencing, and policing data. Existing police data on “high-risk” communities and individuals has been influenced by implicit bias and heavily points to racialized groups. Using this data to train algorithms results in outcomes that criminalise certain demographics. NNs recognise patterns and this pattern recognition becomes racial profiling. This manifests in over-policing of certain areas, exacerbating stereotypes and prejudices, and potentially inciting the same animosity that preventative policing hopes to avoid. Biases are exacerbated while Artificial Intelligence feigns objectivity.
Through advanced pattern recognition alone, Artificial Intelligence represents progress and promise. But progress is not always positive. As we continue to invest in this technology, we must consider the subsequent growth in artificial appetite and exercise care in accommodating it. Artificial Intelligence requires robust data. It is not easy to acquire valid quality data or to certify that our data does not exacerbate stereotypes, but the work is well worth it. Properly satiating AI escapes cycles of garbage, opens up worlds of possibility.
References:
Dr. Rosemary Lee – Who Arted: Weekly Art History (PODCAST)
https://en.wikipedia.org/wiki/Garbage_in,_garbage_out
https://www.geeksforgeeks.org/neural-networks-a-beginners-guide
https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414
https://a16z.com/the-techno-optimist-manifesto
