Computer-based neural networks can be therapeutic as well as fun.
Computer-based neural networks can be therapeutic as well as fun.
going to build computer. Or rather, once we’ve saved enough for the parts, my son and I are going to. This plan received a wee push from a conversation the lad had with my friend Chris. Chris has a computer that runs not one, but two Nvidia GeForce GTX 1080i Ti graphics cards. I know, your mind is blown.
Hardware like this allows you to play beautiful computer games, or to mine Bitcoin, but Chris’s team use them to run neural networks. These are systems designed to mimic the way human brains work – interconnected nodes that do such things as pass on a signal to other nodes when they’re activated, or are designed to do something specific.
Neural networks can “learn” to do things by simulating outcomes, and pruning pathways that don’t contribute to that outcome. In 1992, when I was a student, I built a simple one that learnt how to play noughts and crosses, for example.
About four years after I made my puny neural net, a computer called Deep Blue beat a human chess world champion, Garry Kasparov, for the first time. Things have moved on to a point where last December, AlphaZero, a program developed by DeepMind, which is owned by the same company, Alphabet, that owns Google, beat world computer-chess champion Stockfish 8. AlphaZero taught itself to play chess in about four hours.
Unlike Deep Blue, however, AlphaZero doesn’t always play in a way that looks human. On occasion, it throws away pawns with abandon, yet manages to win. It doesn’t go for the standard openings experts have come to know and love, and it wins.
In short, it wouldn’t pass a Turing test, devised by English mathematician Alan Turing in the 1950s, which requires behaving in a way that a person wouldn’t be able to distinguish from natural human behaviour. Another challenge, for the Loebner Prize, first offered in 1990, subjects artificial intelligence systems to the same test.
One of the early contenders in this “imitation game”, as Turing called it, was Eliza, a program written in the 1960s that responded to typed input in a manner modelled on a Rogerian therapist. Carl Rogers was the influential proponent of a non-directive “client-centred” form of therapy in which the job of the therapist is to reflect back the client’s understandings of their world to help them gain insight into their own problems. Thus, Eliza would take the typed input and reflect it back to the inputter.
So, just as a client of Rogers might say, “I’m feeling frustrated right now”, and have that reflected back as something like, “What I’m hearing is that you’re feeling quite frustrated”, Eliza would parse out the relevant bits of the input and flip it around.
Rogers was also quite important for his recommendations about researching the effectiveness of psychotherapy. At the time, not all psychotherapies – Freudian psychoanalysis, for example – were assessable.
Freud’s ideas were so reliant on the influences of a not-directly accessible unconscious that it was quite difficult to determine how therapeutic improvements might be happening. Rogers eschewed this nebulous unconscious and focused instead on the “fit” between how a client sees themselves in the here and now and how they wish or desire themselves to be.
Therefore, if you feel that therapy is helping you deal with your problems, this should be reflected in a smaller gap between these “actual” and “ideal” selves than was the case at the start of therapy.
I can’t find any indication of what Rogers thought of Eliza, but you can try one of the Elizas on the internet. Don’t give it your bank account if it asks, though. There might be a Nigerian prince on the other end.
Four years after I made my puny “noughts and crosses” neural net, a computer beat a human chess world champion.