基于上下游脑区深度学习模型得到抑郁症、强迫症、选择困难症和偏见的猜想
The conjectures of Depressive disorder，Obsessive-compulsive disorder, Difficult decisions and Prejudice based on the Deep learning model for upstream and downstream brain regions
参考DSM-5的强迫症、选择困难症的相关文字，尝试用上下游脑区深度学习模型进行模拟。我们补充选择困难症一种新情况，前向传播可能在中途返回上游脑区的前额叶。对于偏见的深度学习模型，不理想的数据可能跳过了前额叶；但喜欢的数据经过前额叶后，会跳过带有情绪记忆的脑区。抑郁症的模型可能和情绪记忆、突触兴奋性和神经元活性有关。
首先我回顾提出的上下游脑区的深度模型。这是抑郁症、强迫症、选择困难症和偏见的基本模型。上游脑区的简单信号具有较少的情绪记忆，具有高阶优化的反向传播，更新突触位置和连接的步长更加精细。而下游脑区积累的复杂信号具有更多的情绪记忆，具有低阶优化的反向传播，调整突触位置和连接的步长更加大。
Referring to the DSM-5 Obsessive-compulsive disorder and Difficult decisions, try to simulate it with the Deep learning model for upstream and downstream brain regions. We supplement the new case of Difficult decisions, and the forward propagation may return to the prefrontal lobe of upstream brain regions at halfway. For Prejudice Deep learning model, undesirable data may skip the prefrontal lobe; But favorite data passes through the prefrontal lobe and skips brain regions with emotional memories. Model of Depressive disorder may be related to emotional memory, synaptic excitatory, and neuronal activity.
First of all, I review the proposed Deep learning of upstream and downstream brain regions. This is the basic model for Depressive disorder, Obsessive-compulsive disorder, Difficult decisions, and Prejudice. Where simple signals from upstream brain regions have less emotional memories, have high-order optimized back propagation, and have more refined steps to update synaptic positions and connections. However, the complex signals accumulated in the downstream brain regions have more emotional memories, have low-order optimized back propagation, and the step of adjusting synaptic positions and connections is larger.
1 抑郁症的模型
1 The model of Depressive disorder
我们猜测由于情绪持续低落、心理功能下降、兴趣和乐趣丧失导致脑区的情绪记忆超过了临界值，进而影响突触和神经元。不管是上游脑区还是下游脑区都充满了情绪记忆，另外上游脑区过多的情绪记忆可能和大脑老化、海马体萎缩和海马体硬化有关。由于脑区过多情绪记忆进而影响突触兴奋性和神经元活性，因而改变了大脑的可塑性。
受到单胺假说和神经营养假说的启发，考虑突触范围权重和连接权重的梯度更新对应的学习率，由于突触兴奋性和神经元活性减弱使得学习率数值特别小。
见图1的抑郁症大脑情况，对比图7的正常的大脑情况，较远距离的突触连接已经消失，近距离突触连接即使存在也是弱连接。
We hypothesize that the emotional memories of brain regions exceed the critical value due to by persistent low mood, declining of mental function, loss of interest and enjoyment, which in turn affects synapses and neurons. Both upstream and downstream brain regions are full of emotional memories, and excessive emotional memories in upstream brain regions may be related to brain aging, hippocampal atrophy, and hippocampal sclerosis. Excessive emotional memories in brain regions affect synaptic excitatory, and neuronal activity, so altering brain plasticity.
Inspired by the monoamine hypothesis and the neurotrophic hypothesis, consider the learning rate in gradient of synaptic range weights and connection weights, this learning rate is particularly small due to the weakening of synaptic excitatory, and neuronal activity. See Fig. 1 for the brain regions of Depressive disorder, compared to the normal brain in Fig. 7, where distant synaptic connections have disappeared, and close synaptic connections are weak even if they are still existed.
Fig.1 Brain regions of Depressive disorder
2 强迫症模型
2 The model of Obsessive-compulsive disorder
我们猜想强迫症的模型有两种情况，都和前额叶皮层有关。
一种是过于相信自己较好的经历和急躁，使得前额叶的突触连接权重和范围权重的前向计算陷入局部最优，见图2和8。图2的权重更新参考了局部优权重，较少参考差的权重很容易陷入局部最优。
另一种是童年受到的心灵创伤和不幸经历，使得前额叶的前向计算结果不好，见图3和8。图3的权重更新参考了局部差权重，较少参考优的权重使得计算结果不好。
图8前额叶皮层过于薄。上游皮层的较大计算误差累积到下游皮层需要更多的情绪才能跳出局部最优，进而产生更多的焦虑。
We guess that the model of obsessive-compulsive disorder has two conditions, both related to the prefrontal cortex.
One is to believe too much in their own good experience and impatience, so that the forward propagation of synaptic connection and range weights of the prefrontal lobe falls into local optimum, see Fig. 2 and 8. The weight update in Fig. 2 refers to the local good weights, and with less reference inferior weights can easily fall into the local optimal.
The other is the trauma and unfortunate experience of childhood, which makes the forward propagation of the prefrontal lobe poor, as shown in Fig. 3 and 8. The weight update in Fig. 3 refers to the local inferior weights, and with less reference to the good weights make the calculation results poor.
The prefrontal cortex is too thin in Fig. 8. Too large computational errors in the upstream cortexes accumulate to the downstream cortexes and require more emotions to jump out of the local optimum, which in turn produces more anxiety.
3 选择困难症模型
3 The model of Difficult decisions
而选择困难症是另一种相反情况，也和前额叶皮层有关。
考虑过多的不利情况、有利情况、约束条件和期待过高，特别是不利情况使得上游皮层权重范围较广，搜索效率低也使得上游皮层前向计算不好，见图4和9。
图9前额叶皮层过于厚。上游区更多皮层的海森矩阵进行反向传播，使得算法空间复杂度增多。上游皮层的较大计算误差累积到下游皮层需要更多的情绪才能跳出局部最优，进而产生更多的焦虑。
图5和7是正常的前额叶神经元分布及其皮层厚度。图6和7是更理智的神经元分布及其皮层厚度。
我们考虑一种新的情况，由于选择困难，前向计算的中途返回上游脑区的前额叶，局部陷入死循环。一般而言，前向计算的中途将会到达下游脑区，见图9的紫色箭头。
The Difficult decisions is another opposite situation, also related to the prefrontal cortex.
Considering too many unfavorable situations, favorable situations, the constraints and expectations is too high, especially the unfavorable situations make the range of upstream cortexes weights wider, and the low search efficiency also makes the upstream cortexes forward propagation poor, see Fig. 4 and 9.
The prefrontal cortex is too thick in Fig. 9. The Heisen matrix with more cortexes in the upstream regions is calculated by back propagation, which increases the space complexity of the algorithm. Too large computational errors in the upper cortex accumulate to the downstream cortexes and require more emotion to jump out of the local optimum, which in turn generates more anxiety.
Fig. 5 and 7 show the normal prefrontal neurons distribution and their cortical thickness. Fig. 5 and 6 show a more rational distribution of neurons and their cortical thickness.
We consider a new situation where, due to the Difficult decisions, the forward propagation returns to the prefrontal lobe of the upstream brain regions at halfway, locally caught in an endless loop. In general, the forward propagation will reach to the downstream brain regions at halfway, as shown in Fig. 9.
Fig.2 Distribution of neurons at OCD situation 1, past good experience and impatience
Fig.3 Distribution of neurons at OCD situation 2, childhood misfortune
Fig.4 Distribution of neurons at Difficult decisions
Fig.5 Normal IQ
Fig.6 High IQ
Fig.7 Normal brain regions
Fig.8 Brain regions of OCD
(a) The prefrontal cortex is too thick
(b) The forward propagation returns to the prefrontal lobe
Fig.9 Brain regions of Difficult decisions
4 偏见的模型
4 The model of Prejudice
对于偏见的深度学习模型，不理想的数据可能跳过了前额叶，到达带有情绪记忆的脑区；但喜欢的数据经过前额叶后，跳过带有情绪记忆的脑区。一种是低估目标函数因为缺乏高阶优化简单信号没有很快收敛；一种是高估目标函数因为缺乏情绪记忆之后复杂信号无法跳出局部最优。两种情况见图10，不理想的数据和喜欢的数据的前向计算路径见图中的紫色箭头。低估目标函数时，不同脑区情绪记忆可能出现了骤升，而高估目标时，不同脑区情绪记忆可能出现了骤降。
For Prejudice Deep learning model, undesirable data may skip the prefrontal lobe and reach brain regions with emotional memories; But the favorite data passes through the prefrontal lobe, skipping the brain region with emotional memory. One is to underestimate the objective function because of the lack of higher-order optimization, and the simple signals do not converge quickly. One is to overestimate the objective function, because of lack of emotional memory, complex signals cannot jump out of the local optimum. Both cases are shown in Fig. 10, the forward propagation path for undesirable data and favorite data are shown in the purple arrows in the figure. When the objective function is underestimated, emotional memories in different brain regions may increase sharply, while the objective function is overestimated, emotional memories in different brain regions may decrease sharply.
(a) Undesirable data
(b) Favorite data
Fig.10 Brain regions of Prejudice
作者：陶俊波， 哈工大在读博士。
