February 22nd 2019: Radio Free Mobile updates its coverage of artificial intelligence with the publication of: Reality Bytes: AI: Winter is coming.

RFM research subscribers will receive their copy directly by email. 

Winter is coming. AI and deep learning have delivered great things but only under very controlled conditions. The main limitation of deep learning is that it has no understanding of the task it is performing and is simply matching statistical characteristics to outcomes. This means that it can neither deal with changes in the environment nor new situations. This means that it will fail to live up to most of the hype. The result will be disappointment and falling investment: i.e. a 3^rd AI Winter.

• The 2^nd industrial revolution may end up being more like the 1^st than many would like. Steam power was properly invented in 1698 and took 140 years to change the world. So far, AI is following the same time-line and the emerging challenges may see it continue to do so.
• Hype. RFM thinks that the biggest problem facing AI is the expectations that have been set. Confident predictions of advanced cognitive ability in machines being just around the corner are looking like they will not be fulfilled anytime soon. This has not entered the general consciousness of commentators, consumers, regulators, technology companies or investors as VC investing remains at an all-time high and conference tickets still sell like hotcakes.
• Deep Learning algorithms have no understanding of the functions they perform as they simply match statistical patterns to outcomes. This means that if the data set changes for any reason, the algorithm will often catastrophically fail. In a world of infinite perfectly labelled data and infinite compute power, this would be less of a problem. Unfortunately, this is not reality.
• Compute power. AlphaGo Zero was trained using 300,000x the compute resource compared to AlexNet which set a standard for image recognition in 2012. RFM thinks that AlphaGo Zero does not offer 300,000x the value of AlexNet strongly implying diminishing returns from investments in compute This is significantly exacerbated by the ending of Moore’s Law.
• Perfect data. Deep learning does best in environments which are both finite and stable (e.g. games) as even most sophisticated deep learning algorithms are incapable of extrapolation due to their lack of causal understanding. This is why they struggle in the real world which most of the time is both infinite and unstable. Awareness of this problem is extremely low which has led to confident predictions that the achievements made in finite and stable environments will soon be replicated in the real world.
• AI Winter. The limitations of deep learning are likely to result in it falling far short of the expectations of commentators, technology companies and investors. The result will be disappointment, disillusionment and falling investments. This is exactly what has characterised the 1^st and 2^nd AI winters. RFM sees no reason why a 3^rd would be any different.